Retrospective

Run a structured retrospective after shipping.

Purpose

Objectively verify the entire phase process. Check both output quality AND process compliance.

Input

$ARGUMENTS (optional): Branch name to audit. If empty, uses current branch.

Execution Flow

Step 1: Determine Scope

git branch --show-current
git log --oneline $(git merge-base HEAD <base>)..HEAD
git diff --stat $(git merge-base HEAD <base>)...HEAD

Identify the base branch (from PR target or parent feature branch).

Step 2: Spawn 2 Agents in Parallel

Spawn 2 agents using the Task tool (subagent_type: "general-purpose", model: "<choose per task>") in parallel.

MANDATORY: Always specify an explicit model parameter. Choose the appropriate model based on task complexity (haiku for lightweight, sonnet for standard, opus for complex reasoning). Never omit model (default inherit may fail in parallel spawning).

Auditor and Researcher are independent — no inter-agent communication needed.

Agent 1: Auditor

5-principle compliance verification. For detailed prompt, read ${CLAUDE_PLUGIN_ROOT}/skills/retrospective/references/auditor-prompt.md.

Output: PASS/PARTIAL/FAIL per principle with evidence.

Agent 2: Researcher

Code quality and architecture analysis. For detailed prompt, read ${CLAUDE_PLUGIN_ROOT}/skills/retrospective/references/researcher-prompt.md.

Output: Strengths, Weaknesses, Recommendations, Metrics.

Step 3: Integrate Reports

Integrate both agent reports:

1. Audit result fixes: If PARTIAL/FAIL found, apply fixes as Fixer
2. workflow-recording.md update: Add phase metrics
3. Learnings optimization: Read and optimize docs/dev-cycle-learnings.md (see Learnings PDCA below)
4. MEMORY.md update: Record process lessons

Step 4: Fix (if needed)

Based on Auditor/Researcher findings:

• Code issues: fix + test + commit
• Skill-level issues: Output a GitHub Issue suggestion (see Learnings PDCA below) — do NOT modify cached SKILL.md files
• Process issues: Record lessons in MEMORY.md / CLAUDE.md

After fixes: code:review-commit via Skill tool + approve-review.sh + commit.

Step 5: Commit Retrospective Artifacts

docs(retro): Phase N retrospective findings and improvements

Step 6: Push Updated Branch

git push origin <branch>

Output Report

## Retrospective Complete

### Audit Summary
| Principle | Status | Notes |
|-----------|--------|-------|
| DDD/TDD/DRY/ISSUE/PROCESS | pass/partial/fail | ... |

Score: X/5

### Code Quality
- Strengths / Weaknesses / Recommendations

### Metrics
- New files, tests, coverage, review iterations

### Improvements Applied
- Learnings updated (N new, N merged, N resolved), memory updated, code fixes

Output Rules

For shared output language rules, read ${CLAUDE_PLUGIN_ROOT}/skills/_shared/output-rules.md.

Important Notes

• This skill runs autonomously — no user confirmation between steps
• Agents must be brutally honest — no flattery, no sugar-coating
• Both technical issues AND process compliance must be verified
• Metrics recorded in docs/research/workflow-recording.md
• Process lessons go in project MEMORY.md
• Auditor + Researcher agents run via Task tool in parallel (TeamCreate is unnecessary — agents are independent)
• All user-facing output MUST follow the user's configured language setting — SKILL.md being in English does not change this

Learnings PDCA

For detailed Learnings PDCA procedures (Project-Side optimization + Plugin-Side GitHub Issue suggestion), read ${CLAUDE_PLUGIN_ROOT}/skills/retrospective/references/learnings-pdca.md.